Update error message when attempting inference on model with 0 workers about serve HOT 4 CLOSED

@chauhang

User can provide the initial-workers while registering the models through management API.

As documented in management api doc :

User may want to create workers while register, creating initial workers may take some time, user can choose between synchronous or synchronous call to make sure initial workers are created properly.

The asynchronous call will return before trying to create workers with HTTP code 202:

curl -v -X POST "http://localhost:8081/models?initial_workers=1&synchronous=false&url=https://<s3_path>/squeezenet_v1.1.mar"

< HTTP/1.1 202 Accepted
< content-type: application/json
< x-request-id: 29cde8a4-898e-48df-afef-f1a827a3cbc2
< content-length: 33
< connection: keep-alive
< 
{
  "status": "Worker updated"
}

The synchronous call will return after all workers has be adjusted with HTTP code 200.

curl -v -X POST "http://localhost:8081/models?initial_workers=1&synchronous=true&url=https://<s3_path>/squeezenet_v1.1.mar"

< HTTP/1.1 200 OK
< content-type: application/json
< x-request-id: c4b2804e-42b1-4d6f-9e8f-1e8901fc2c6c
< content-length: 32
< connection: keep-alive
< 
{
  "status": "Worker scaled"
}

We are planning to update the error message as follows :

{
"code": 503,
"type": "ServiceUnavailableException",
"message": "No worker available to serve request for model <model_name>. Please use scale workers api to add workers."
}

Please let us know your thoughts.

from serve.

Changing the Error Message sounds like a reasonable way. Should we set this to a default value of 1 so clients can make inference?

from serve.

Leaving the default workers as 0 is the safest option, and the user still has the option of specifying initial_workers when registering the model. I would favor leaving the default as it is.

The two measures I'd recommend:

• Change the 503 error message, as @mycpuorg suggested
• Include a warning in the response to the call that registers the model, reminding the user that they have no workers dedicated to the model.

With this added communication to mitigate user surprise, I think we'd be in good shape.

from serve.

I ran into a similar issue when following an example here:

curl -X POST "localhost:8081/models?model_name=resnet152&url=resnet-152-batch.mar&batch_size=4&max_batch_delay=5000&initial_workers=3&synchronous=true"

I plugged in my own model. Is it an outdated piece?

from serve.